Medicine Is Messy. That Is Where the Art Lives.
AI Can Clean the Canvas. Should It Finish the Picture?
The Future of Clinical Decision Support is a new series from MDCalc exploring how AI, evidence, and physician judgment are reshaping medicine. These essays are intended to spark discussion about where technology is taking clinical decision support – and where physicians feel it should go.
I can put two patients with remarkably similar MRIs in two different exam rooms and recommend surgery to one and continued conservative care to the other. Both may have similar lumbar stenosis, symptoms, neurologic exams, and prior treatments. If you reduced them to imaging, diagnosis, and a handful of clinical variables, they might look almost identical.
One patient may tell me his goal is to keep golfing, traveling, and playing with his grandchildren. He understands the risks of surgery, but his current quality of life is unacceptable. Another patient with nearly the same MRI may be caring for a spouse, functioning reasonably well, and unwilling to risk losing independence during recovery. A third may desperately want surgery, but after talking with them I am not convinced the problem they want fixed is actually coming from their MRI.
That is what I kept thinking about while reading Graham Walker’s essay, Medicine Was Never Clean. His point about medicine being built on imperfect evidence and noisy inputs is important. But I think there is another kind of messiness we need to separate from the quality of those inputs. Even if AI someday gives us remarkably clean information, the clinical decision itself may still be messy.
There is plenty of mess in medicine I would be thrilled to eliminate. The operative report buried in another health system, the medication list nobody reconciled, therapy notes scattered across a chart, missing outside imaging, and twenty minutes spent reconstructing what happened to a patient over the last three years are not the art of medicine. They are noise around medicine, and I want AI to attack that noise aggressively.
AI should be able to assemble the relevant history before I enter the room, understand what treatments have been attempted, identify inconsistencies, surface important imaging, summarize the evidence, and show me where that evidence becomes weaker for the particular patient in front of me. It may eventually become far better than physicians at processing that amount of information. That does not threaten me as a physician. I want that help.
Now imagine AI becomes extraordinarily good at predicting outcomes. It tells me a patient has a 72 percent probability of meaningful functional improvement after surgery, a 19 percent probability of little meaningful change, and a clearly quantified risk of complications. It reaches those numbers using imaging, medical history, functional data, patient goals, and outcomes from hundreds of thousands of similar patients.
I would love to have that information. But it still does not necessarily answer whether that patient should have surgery. Is a 72 percent chance of improvement worth the risk to this person? What does meaningful improvement mean to them? What are they willing to tolerate to get there? What would a complication mean in the context of their life?
AI can estimate what is likely to happen. The moment it decides what should happen, it is also applying values about which outcomes matter and which risks are acceptable. Prediction and judgment overlap, but they are not interchangeable.
That does not give physicians permission to hide behind experience whenever evidence disagrees with us. We are biased. We anchor on diagnoses and overvalue our own experiences. AI may become one of the best tools we have for challenging those instincts and showing us what we missed.
The art of medicine should never become an excuse to ignore evidence. It is the work that remains after the evidence has been made as clear as possible. That is what good clinical AI looks like to me. Clean up the information surrounding the encounter. Reduce the noise between encounters. Organize the patient’s story, expose what we missed, quantify risk where possible, and make uncertainty more visible rather than hiding it. Then give physicians and patients the capacity to work through what that information actually means for the person sitting in the room.
Physicians, patients, and the people building these systems need to define that boundary together. Where is the mess simply noise that technology should eliminate? Where can better prediction improve our judgment? And where does the remaining uncertainty become part of the work that physicians and patients still need to navigate together?
Cleaner information will not always produce a clean clinical decision. Sometimes what remains is uncertainty, competing values, imperfect choices, and two people trying to decide which risk is worth taking.
Which parts of the canvas do we want AI to clean up, and which parts of the picture do physicians and patients still need to finish together?
Because some of the mess in medicine is a problem. Some of it is where the art lives.
Clay Dorenkamp, D.O.
Orthopedic Spine Surgeon
Michigan Orthopedic Center